[maleadt]

The default method_table_view nests a StackedMethodTable per method table, and GPUInterpreter is parametric in the view, so back-ends that stack a different number of tables get different interpreter types. The inference code specialized on that type is only shared between packages whose stacks happen to have the same depth. GPUCompiler's own precompilation workload (native target, one table) covers none of the back-ends.

This showed up as a Metal.jl precompilation regression (32 → 38.7 s on CI, JuliaGPU/Metal.jl#1129). Metal stacks three tables plus SHARED_METHOD_TABLE. Until KernelAbstractions 0.10.5, so did KA's POCL back-end, whose precompilation workload launches a kernel, so Metal reused the inference code in KA's package image. JuliaGPU/KernelAbstractions.jl#944 added a fourth table to POCL, and Metal had to compile it all again.

This PR replaces the nested chain with a flat MethodTableStack that holds the tables in a vector. It has the same lookup semantics, so every back-end that returns its tables from method_tables (and the native target) shares one interpreter type, and with it the inference code GPUCompiler precompiles.

Measured on an M1, Julia 1.13, user CPU time of Base.compilecache (2 interleaved rounds):

	2.21	this PR
Metal.jl, with KA 0.10.5	69.4 / 67.6 s	54.2 / 53.3 s
Metal.jl, with KA's workload disabled	75.6 / 73.0 s	58.0 / 59.7 s
GPUCompiler	36.8 / 37.2 s	36.9 / 37.1 s

Compiling 29 new Metal kernels in a warm session takes the same time (1.23–1.32 s), so lookups through the flat stack don't slow down inference.

This doesn't change the back-end interface. StackedMethodTable is still available, and back-ends that override method_table_view keep working, but they get their own interpreter type and don't share the precompiled code. OpenCL.jl does that today and could switch to method_tables, though that would also add SHARED_METHOD_TABLE to its stack.

[codecov]

Codecov Report

❌ Patch coverage is 86.95652% with 6 lines in your changes missing coverage. Please review.
✅ Project coverage is 92.81%. Comparing base (7660c58) to head (6eb6bac).

Files with missing lines	Patch %	Lines
src/jlgen.jl	86.95%	6 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1080      +/-   ##
==========================================
- Coverage   92.84%   92.81%   -0.04%     
==========================================
  Files          31       31              
  Lines        7732     7775      +43     
==========================================
+ Hits         7179     7216      +37     
- Misses        553      559       +6     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.